Self-interference suppression method and device for full-duplex communication

Abstract

The invention discloses a self-interference suppression method and device for full-duplex communication in the technical field of full-duplex communication. The self-interference suppression method comprises the following steps: S1, acquiring an input signal and preprocessing the input signal; s2, performing CZT frequency domain conversion on the input signal, and outputting a complex frequency spectrum signal cm (n); s3, inputting the complex frequency spectrum signal cm (n) into an ML-CVNN multilayer complex valued neural network, and extracting a channel characteristic # imgabs0 #; S4, inputting the channel characteristic # imgabs1 # into a CLDNN network for channel prediction, and obtaining a predicted self-interference signal; and S5, subtracting the predicted self-interference signal from the received signal, and outputting a received signal rout (n). According to the self-interference suppression method, complex nonlinear characteristics are accurately modeled through a complex valued neural network, and the amplitude and phase characteristics of signals are more effectively captured; through the CZT-ML-CVNN architecture, the dynamic channel prediction precision and the spectral resolution are improved, and the power spectrum leakage is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of full-duplex communication, and in particular to a method and device for suppressing self-interference of full-duplex communication. Background Art

[0002] Full-duplex communication allows the transmitter and receiver to work in the same frequency band at the same time, but the self-interference problem introduced by this limits the practical application of full-duplex communication. Traditional self-interference elimination technology mainly relies on the following methods:

[0003] 1) Passive suppression: reduce the power of the transmitted signal coupled to the receiver through antenna design, isolation structure and other means. For example: antenna isolation: increase the physical distance between the transmitting antenna and the receiving antenna; directional antenna: reduce the overlap of the main beams of the transmitted and received signals; shielding material: use absorbing materials between antennas to reduce electromagnetic coupling.

[0004] 2) Analog cancellation: Generate an anti-phase signal identical to the self-interference signal in the analog domain and superimpose it on the received signal to cancel the interference. Such as signal replication and phase adjustment: Use circuits to generate copies of the self-interference signal and match them by adjusting the phase and amplitude; Broadband adaptation: For broadband signals, more accurate interference matching is achieved through segmented processing.

[0005] 3) Digital elimination: Digital self-interference cancellation (DSIC) is the core part, which is used to process the residual self-interference signal after passing through the analog domain. There are many digital elimination methods, such as using linear models: estimating and reconstructing the self-interference signal through known self-interference signals and channel characteristics, and then canceling it in the digital domain; nonlinear models: in order to deal with nonlinear interference introduced by hardware, polynomial models or machine learning methods are often used.

[0006] In the IBFD (In-Band Full-Duplex) system, by using the CLDNN network (CNN, LSTM and DNN fusion), the experimental results show that it is significantly better than the traditional method in eliminating self-interference performance. However, the current digital elimination method has the defects of insufficient dynamic channel adaptability, limited frequency domain accuracy of feature extraction, and limited nonlinear modeling ability of self-interference. Summary of the invention

[0007] In view of this, an embodiment of the present invention provides a method and device for suppressing self-interference in full-duplex communication, so as to solve or partially solve the above problems.

[0008] In a first aspect, an embodiment of the present invention provides a method for suppressing self-interference in full-duplex communication, comprising the following steps:

[0009] S1: Get input signal and preprocess it;

[0010] S2: Perform CZT frequency domain conversion on the input signal and output a complex spectrum signal c m (n);

[0011] S3: The complex spectrum signal c m (n) Input ML-CVNN multi-layer complex-valued neural network to extract channel characteristics

[0012] S4: The channel characteristics Input the CLDNN network to perform channel prediction and obtain the predicted self-interference signal;

[0013] S5: Subtract the predicted self-interference signal from the received signal and output the received signal r out (n).

[0014] According to a specific implementation method of an embodiment of the present invention, in step S1, the input signal includes: a baseband signal x(t) generated by a transmitting end and a received signal r(t) containing a self-interference component; the preprocessing is to digitize x(t) and r(t) to obtain equivalent discrete baseband signals x(n) and r(n).

[0015] According to a specific implementation of the embodiment of the present invention, step S2 specifically includes the following steps:

[0016] S2.1: CZT calculation of Doppler spectrum: Use the CZT conversion formula to convert the input signal r(n) into a frequency domain signal:

[0017] S2.2: Channel parameter extraction: The key frequency domain features are obtained through peak detection as follows: a m : the amplitude of each path; f m : Doppler frequency; φ m : phase shift;

[0018] S2.3: Output result: Output the complex spectrum signal c m (n).

[0019] According to a specific implementation of an embodiment of the present invention, in step S3, the ML-CVNN multi-layer complex-valued neural network extraction process is as follows:

[0020] Input layer: The complex spectrum signal c output by CZT m (n) as input to ML-CVNN;

[0021] Hidden layer: Use the ML-CVNN multi-layer complex-valued neural network for channel modeling:

[0022]

[0023] In the formula, |w lji ∣: the magnitude of the connection weight; θ lji : the phase of the connection weight;

[0024] Activation function:

[0025] Output layer: The output layer generates the predicted channel characteristics

[0026] According to a specific implementation of an embodiment of the present invention, in step S4, the CLDNN network prediction method is as follows:

[0027] Input layer: The channel characteristics extracted by the ML-CVNN multi-layer complex-valued neural network Input into the CLDNN network;

[0028] CNN module: uses convolutional layers to extract frequency domain features;

[0029] LSTM module: uses long short-term memory network to extract time-related features;

[0030] DNN module: maps features to independent space and generates interference signal prediction results;

[0031] Output layer: outputs the predicted self-interference signal.

[0032] According to a specific implementation of the embodiment of the present invention, step S5 specifically includes the following steps:

[0033] S5.1: Signal reconstruction: Reconstructing the self-interference signal according to the predicted self-interference signal

[0034] S5.2: Interference cancellation: Subtract the self-interference signal from the received signal r(n)

[0035]

[0036] S5.3: Output the received signal r after self-interference suppression out (n).

[0037] According to a specific implementation of an embodiment of the present invention, the hardware implementation steps of the self-interference suppression method are as follows:

[0038] a. Get the dataset;

[0039] b. Use the Keras library to train the ML-CVNN multi-layer complex-valued neural network and the CLDNN network, using a mini-batch gradient descent algorithm during the training process;

[0040] c. Deploy the trained ML-CVNN multi-layer complex-valued neural network and the CLDNN network on the FPGA to process the received signal in real time;

[0041] d. Compare the error between the output signal and the ideal signal and calculate the performance index.

[0042] According to a specific implementation of an embodiment of the present invention, the data set comes from a test platform of a full-duplex communication system, and includes a received signal and a self-interference signal. The data in the data set includes a time series, spectrum information, and an annotation of the received signal.

[0043] In a second aspect, an embodiment of the present invention provides a self-interference suppression device for full-duplex communication, comprising:

[0044] An acquisition module is used to acquire input signals and perform preprocessing;

[0045] A conversion module is used to perform CZT frequency domain conversion on the input signal and output a complex spectrum signal c m (n);

[0046] An extraction module is used to extract the complex spectrum signal c m (n) Input ML-CVNN multi-layer complex-valued neural network to extract channel characteristics

[0047] A prediction module is used to convert the channel characteristics Input the CLDNN network to perform channel prediction and obtain the predicted self-interference signal;

[0048] An output module is used to subtract the predicted self-interference signal from the received signal and output the received signal r out (n).

[0049] The embodiments of the present application have at least the following beneficial effects:

[0050] 1. This self-interference suppression method introduces a complex-valued neural network (CVNN), especially in combination with the Chirp Z transform (CZT), to accurately extract time-frequency domain features of the signal, improve the ability to capture dynamic changes of the channel, and provide CLDNN with better feature input.

[0051] 2. This self-interference suppression method uses CZT instead of traditional FFT, which significantly improves the resolution of spectral details and provides more accurate frequency domain input for the CNN layer of CLDNN.

[0052] 3. This self-interference suppression method introduces ML-CVNN as a nonlinear feature extraction module at the front end of the CLDNN network. It captures the amplitude and phase relationship of the signal more efficiently through the complex-valued neural network, reduces the learning difficulty of subsequent networks, and improves the effect of self-interference modeling. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following is a brief introduction to the drawings required for the specific embodiments or the description of the prior art. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn according to the actual scale.

[0054] Figure 1 A flow chart of a method for suppressing self-interference in full-duplex communication provided by an embodiment of the present invention is shown;

[0055] Figure 2 A schematic diagram of an equivalent baseband digital signal model in an embodiment of the present invention is shown;

[0056] Figure 3 A schematic diagram of a complex-valued neural network architecture in an embodiment of the present invention is shown;

[0057] Figure 4 A schematic diagram of the structure of the ML-CVNN network in an embodiment of the present invention is shown;

[0058] Figure 5 A principle block diagram of a method for suppressing self-interference in an embodiment of the present invention is shown;

[0059] Figure 6 The structure block diagram of a self-interference suppression device for full-duplex communication provided by an embodiment of the present invention is shown. DETAILED DESCRIPTION

[0060] The following embodiments of the technical solution of the present invention are described in detail in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and are therefore only used as examples, and cannot be used to limit the protection scope of the present invention.

[0061] It should be noted that, unless otherwise specified, the technical terms or scientific terms used in this application should have the common meanings understood by those skilled in the art to which the present invention belongs.

[0062] Figure 1 A flowchart of a method for suppressing self-interference in full-duplex communication provided by an embodiment of the present invention is shown in FIG. Figure 1 , the method comprises the following steps:

[0063] S1: Get input signal and preprocess it;

[0064] The input signals include:

[0065] Time domain signal: baseband signal x(t) generated by the transmitter;

[0066] Received signal: The received signal r(t) containing the self-interference component.

[0067] Preprocessing refers to digitizing x(t) and r(t) to obtain equivalent discrete baseband signals x(n) and r(n). Data normalization and denoising ensure the stability and accuracy of the signal when it is input into the model.

[0068] Specifically, the equivalent baseband digital signal model is as follows: Figure 2 shown.

[0069] See also Figure 2 The system has N subcarriers, X(k) (k = 0, 1, ... N-1) is the frequency domain data, and the time domain signal is obtained through inverse fast fourier transform (IFFT). The continuous signal x(t) is obtained through the digital-to-analog converter (DAC) module, and then the RF output signal x is obtained through up-conversion and PA amplifier. PA (t). The PA model adopts the widely used memory polynomial model, and the equivalent baseband digital signal of the RF output is:

[0070]

[0071] Among them, x(n) is the oversampled digital signal, K is the order of nonlinearity, K is the highest order, and since only odd-numbered high-order signals will affect the system, k is an odd number; Q is the corresponding memory depth, and the maximum memory depth is Q-1; a k,q The representative order is k, and the memory depth is the coefficient corresponding to the q signal. From formula (1), it can be seen that PA not only produces nonlinear effects, but also exhibits memory characteristics.

[0072] RF signal x PA (n) Through the self-interference channel h st (t) reaches the local receiving end, let h(t) = h st (t)-h′(t) represents the joint equivalent channel of the actual self-interference channel and the analog domain cancellation channel. Ignoring other cancellation methods, the digital signal at the receiving end is:

[0073] r(t)=x PA (t)*h(t) (2);

[0074] Where * represents the convolution operation. After the signal is down-converted and ADC module, it becomes the baseband digital signal r(n). Substituting equation (1) into equation (2), the expression of r(n) is:

[0075]

[0076] It is assumed that the equivalent channel h has L paths in total, and h(l) is the attenuation coefficient of the lth path. Further, equation (3) can be rearranged as:

[0077]

[0078] in, is the coefficient corresponding to the nonlinear memory signal with k-order memory depth p.

[0079] Since the local receiving end in the full-duplex system knows the transmitted signal x(n), it can be known from equation (4) that the baseband signal r(n) at the receiving end is the coefficient b to be estimated. k, The coefficients are estimated by the algorithm, and the self-interference signal is reconstructed, which is then subtracted from the received signal to suppress the self-interference signal.

[0080] S2: Perform CZT frequency domain conversion on the input signal and output a complex spectrum signal c m (t); comprising the following steps:

[0081] S2.1: CZT calculation of Doppler spectrum: Use the CZT conversion formula to convert the time domain signal r(n) into a frequency domain signal:

[0082]

[0083] CZT provides higher frequency resolution than traditional FFT, especially in narrow frequency bands.

[0084] S2.2: Channel parameter extraction: Obtain key frequency domain features through peak detection: a m : the amplitude of each path; f m : Doppler frequency; φ m : Phase offset.

[0085] S2.3: Output result: Output complex spectrum signal c m (n), c m The spectral features of (n) are used as the input of the subsequent network.

[0086] The CZT transformation is described as follows:

[0087] (1) Channel model:

[0088] The fast fading channel is modeled as the superposition of multiple sine waves, that is, the complex signal caused by multipath reflection is expressed as:

[0089]

[0090] Among them, a m : The amplitude of each path; f m : Doppler frequency; φ m : Phase offset.

[0091] The model assumes that the above parameters remain constant or change slowly over short periods of time.

[0092] (2) The key steps of the CZT method include:

[0093] 1) Doppler spectrum estimation:

[0094] CZT is used to calculate the Doppler spectrum of the channel, and its advantages are:

[0095] Reduce frequency quantization error through interpolation; reduce computational cost compared to zero-padded FFT (ZP-FFT); CZT effectively refines spectrum details by focusing on the area close to zero frequency in the spectrum, avoiding the computational burden brought by full-band calculations.

[0096] 2) Channel parameter estimation:

[0097] Apply peak detection to the calculated Doppler amplitude spectrum: the amplitude peak determines the path amplitude Corresponding frequency determines Doppler frequency The phase shift is determined by the value of the phase spectrum at the corresponding frequency.

[0098] Predict the future channel response based on these parameters:

[0099]

[0100] 3) Window function processing: A Hanning window is applied to the observed channel response before calculating the CZT to reduce spectral leakage.

[0101] (3) Advantages:

[0102] The CZT method significantly improves the spectral resolution (from 14 Hz in NZP-FFT to 1.9 Hz).

[0103] For channel power and phase prediction, the error of the CZT method (2 dB for power and 0.5 radian for phase) is much smaller than that of the ZP-FFT and non-prediction methods.

[0104] Channel prediction based on Doppler spectrum effectively disperses the noise power, so it is more robust to noise when the observation frequency band is narrow.

[0105] The subsequent joint model of CZT and CVNN first uses CZT to perform time-frequency domain transformation on the received signal to extract more refined frequency domain features, which has higher frequency resolution than the traditional fast Fourier transform (FFT); then the output of CZT is used as input and combined with the complex-valued multilayer perceptron (CZT-ML-CVNN) to predict channel fading characteristics and obtain more accurate features.

[0106] S3: The complex spectrum signal c m (n) Input ML-CVNN multi-layer complex-valued neural network to extract channel characteristics

[0107] Among them, the input layer: the complex spectrum signal c output by CZT m (t) is used as the input of ML-CVNN. The input layer receives c m (n-1), c m (n-2),…,c m (n-IML) time series.

[0108] Hidden layer: Channel modeling using a multi-layer complex-valued neural network:

[0109]

[0110] In the formula, |w lji ∣: the magnitude of the connection weight; θ lji : The phase of the connection weights.

[0111] Activation function:

[0112] Modeling nonlinear channel characteristics through amplitude activation and phase preservation.

[0113] Output layer: The output layer generates predicted channel characteristics Used for subsequent CLDNN model training.

[0114] What needs to be explained about complex-valued neural networks is:

[0115] 1. Basic principles:

[0116] CVNNs can process complex signals (including both amplitude and phase), improving the traditional method of decomposing complex numbers into real and imaginary numbers.

[0117] The characteristic of CVNNs is that they directly model the phase rotation and amplitude modulation of the signal through operations in the complex domain, thereby retaining the essential characteristics of the signal, improving the accuracy and efficiency of signal processing, and learning nonlinear relationships based on the characteristics of the input signal to accurately model the interference signal. The error feedback between its output and the target interference signal is used to update the weights and optimize the network performance.

[0118] 2. Unique advantages of CVNN:

[0119] Efficient processing of complex-valued signals: CVNN can directly capture the phase and amplitude changes of the signal, which is particularly important in modeling self-interference signals.

[0120] Modeling of phase rotation and amplitude attenuation: Through complex-valued operations, CVNN can more effectively adjust for phase mismatches and accurately model the nonlinear characteristics of signals.

[0121] Reduce degrees of freedom: The operational characteristics of complex numbers reduce the redundancy of network parameters, thereby improving training efficiency.

[0122] Dynamic adaptability: CVNN can dynamically adjust network weights based on real-time feedback to adapt to changes in channel conditions, thereby maintaining efficient self-interference cancellation performance in dynamic environments. After nonlinear cancellation using the CVNN N network, it is very close to the bottom noise of the receiver.

[0123] 3. The architecture of complex-valued neural networks, such as Figure 3 As shown:

[0124] Layer design: The network consists of input layer, hidden layer and output layer, and all weights are complex values.

[0125] Activation function: Use specially designed complex-valued activation functions, such as the complex-valued Tanh or split-sigmoid functions, to process both the real and imaginary parts of complex numbers.

[0126] Output: The output is the suppressed signal, which is purified by subtracting the self-interference signal.

[0127] 4. Optimization and training:

[0128] Loss function: Mean square error (MSE) or signal-to-noise ratio (SNR) is used as the loss function to measure the quality of the predicted self-interference signal.

[0129] Training method: Optimize network weights based on the complex backpropagation algorithm (CBP).

[0130] 5. Real-time signal processing:

[0131] When deployed, CVNN is used as an adaptive filter under dynamic channel conditions, and the model can be updated in real time to handle complex nonlinear interference. ComplexNN provides a standard module in complex form under the PyTorch framework without any additional trainable parameters. The parameters and calling methods of each module in this framework are consistent with the PyTorch framework, without additional learning cost.

[0132] In this embodiment, the complex-valued neural network is further improved into a new structure suitable for self-interference suppression:

[0133] CZT-ML-CVNN model: ChirpZ transform (CZT) combined with linear extrapolation and Lagrangian extrapolation is used to predict frequency domain parameters, and ML-CVNN multi-layer complex-valued neural network is combined to predict channel fading through time-frequency transform and complex-valued network, which significantly improves the bit error rate (BER) performance.

[0134] The structure of ML-CVNN is divided into three layers, such as Figure 4 As shown:

[0135] Input layer: input signal c m (n-1), c m (n-2),…,c m (n-IML) are passed to the neurons of the hidden layer respectively.

[0136] Hidden layer: Each hidden neuron receives a weighted input signal and generates an intermediate output z1.

[0137] Output layer: final output prediction signal

[0138] The connection weights between neurons are represented by complex values, and each weight w lji Included amplitude |w lji | and phase θ lji This allows the network to directly model the amplitude and phase changes of complex signals, preserving the essential characteristics of the signal.

[0139] The core contents of ML-CVNN are as follows:

[0140] 1) Internal state:

[0141] The input signal of the neuron is weighted, combined with amplitude scaling and phase rotation, to obtain the internal state u of the neuron. lj :

[0142]

[0143] 2) Activation function:

[0144] z lj For the intermediate output of the hidden layer of the neural network, an amplitude-phase type activation function is used:

[0145]

[0146] Amplitude: By tanh(|u lj |) performs nonlinear scaling activation, similar to ReLU in real-valued networks.

[0147] Phase: e jarg(l j) Preserve the phase information of the input signal.

[0148] 3) Prediction output:

[0149] Neurons in the output layer Calculated directly based on the output z1 of the hidden layer.

[0150] 4) Weight update mechanism:

[0151] ML-CVNN uses the gradient descent method to update weights, allowing the network to learn channel characteristics. Through limited iterative updates of weights, the training time is significantly reduced, making it suitable for dynamic channel environments.

[0152] 5) Real-time optimization method:

[0153] Since the channel characteristics are continuous, the weights do not need to converge completely each time they are updated. Based on this idea, a weak regularity update strategy is adopted:

[0154] After stopping the update of each iteration, the current weight is directly used as the initial value for the next update.

[0155] Through continuous learning, the training cost is reduced while dynamically tracking channel changes.

[0156] S4: Channel characteristics Input the CLDNN network to perform channel prediction and obtain the predicted self-interference signal;

[0157] Among them, the CLDNN network includes:

[0158] Input layer: channel features extracted by ML-CVNN Input into CLDNN.

[0159] CNN module: Use the convolutional layer to extract frequency domain characteristics and reduce the fluctuation of frequency domain characteristics.

[0160] LSTM module: Use the long short-term memory network (LSTM) to extract time-related features.

[0161] Model the dynamic changes of the input sequence and predict future channel characteristics.

[0162] DNN module: maps features to an independent space to generate more accurate interference signal prediction results.

[0163] Output layer: finally outputs the predicted self-interference signal.

[0164] Further explanation about the CLDNN network is:

[0165] From the above equivalent digital signal model, it can be seen that the self-interference of the RF device will produce a nonlinear part of the memory length, so the task of digital interference elimination is closely related to time-related problems.

[0166] CLDNN was first used for speech recognition tasks. It combines the advantages of CNN, Long Short-Term Memory (LSTM) and Deep Neural Network (DNN). Compared with the polynomial method, it is easier to reconstruct nonlinear self-interference by inputting training samples into the neural network for training. In addition, self-interference with high-dimensional features and high nonlinearity is difficult to handle through modeling, and CNN has advantages in solving problems that are difficult to solve with traditional mathematical models. Therefore, applying CLDNN to nonlinear digital cancellation is expected to achieve better performance.

[0167] In terms of modeling capabilities, CNN is good at reducing frequency domain changes, LSTM can provide long-term memory, so it has a wide range of applications in the time domain, and DNN is suitable for mapping features to independent spaces. In CLDNN, CNN, LSTM and DNN are connected and fused into one network to achieve better performance than a single network.

[0168] The general structure of the CLDNN network is that the input layer is time-domain related features, two layers of CNN are connected to reduce frequency domain changes, the output of CNN is fed into several layers of LSTM to reduce time domain changes, and the output of the last layer of LSTM is input into the fully connected DNN layer, the purpose is to map the feature space to the output layer that is easier to classify. Compared with the previous fusion of CNN, LSTM and DNN, which are trained separately and then fused together through the fusion layer, CLDNN trains the three networks at the same time. Experiments have shown that the performance of LSTM will be improved if it inputs better features, and CNN is used to reduce changes in the frequency domain to make LSTM input more adaptive features. Adding DNN increases the depth between the hidden layer and the output layer to obtain stronger prediction capabilities.

[0169] S5: self-interference signal cancellation; subtract the predicted self-interference signal from the received signal and output the received signal r out (n), specifically comprising the following steps:

[0170] S5.1: Signal reconstruction: Reconstruct the self-interference signal based on the prediction results of CLDNN

[0171] S5.2: Interference cancellation: Subtract the predicted self-interference signal from the received signal:

[0172]

[0173] S5.3: Output the received signal r after self-interference suppression out (n).

[0174] The principle block diagram of the self-interference suppression method is as follows Figure 5 As shown. This method follows the idea of ​​CLDNN and transforms the received signal in the time-frequency domain through CZT before entering the network to extract the frequency domain features of the signal. The output of CZT is used as input, combined with the complex-valued multilayer perceptron (CZT-ML-CVNN) to obtain more accurate features, and the obtained features are used as the input of CLDNN for training to obtain stronger prediction capabilities for the nonlinear part of self-interference and high-dimensional features.

[0175] The embodiments of the present invention have the following technical effects:

[0176] 1. CZT conversion:

[0177] The ChirpZ transform (CZT) is used to perform high-resolution frequency domain conversion on the time domain signal and extract key characteristics (such as Doppler frequency, amplitude and phase) in the spectrum.

[0178] Compared with the traditional FFT method, CZT can more accurately obtain frequency domain features within a specific frequency band, and is especially suitable for dynamic channel environments.

[0179] 2. ML-CVNN channel modeling and feature extraction:

[0180] A multi-layer complex-valued neural network (ML-CVNN) is used to model and extract the nonlinear characteristics of the self-interference signal.

[0181] ML-CVNN can directly process complex signals, preserve the amplitude and phase information of the signals, and generate channel features for prediction.

[0182] Complex-valued neural network (CVNN): Accurately model complex nonlinear characteristics; more effectively capture the amplitude and phase characteristics of the signal; reduce the degrees of freedom of training parameters and reduce the risk of overfitting.

[0183] CZT-ML-CVNN: High dynamic channel prediction accuracy; significantly improved spectral resolution, reduced power spectrum leakage; higher computational efficiency compared to traditional FFT.

[0184] CLDNN architecture: combines the advantages of CNN, LSTM and DNN, fully utilizing the time and frequency characteristics of the signal; outperforms traditional methods under dynamic and complex channel conditions.

[0185] 3. CLDNN prediction and classification:

[0186] The features extracted from ML-CVNN are input into CLDNN (a neural network combining CNN, LSTM and DNN) for further training and optimization to predict future channel characteristics.

[0187] CLDNN achieves more accurate interference signal prediction and classification by combining CNN's frequency domain feature extraction, LSTM's time correlation learning, and DNN's feature mapping capabilities.

[0188] 4. Real-time self-interference elimination:

[0189] Finally, the predicted self-interference signal is subtracted from the received signal, and the purified received signal is output.

[0190] The system can dynamically adapt to changes in the channel environment and is suitable for full-duplex communications in high-speed mobile scenarios.

[0191] 5. Wide range of application scenarios, such as full-duplex communication base stations in 5G and future 6G communication systems; dynamic self-interference suppression in radar systems; satellite communications in high-dynamic scenarios.

[0192] The hardware implementation steps of the embodiment of the present invention are as follows:

[0193] a. Dataset and training:

[0194] Dataset: The dataset in the code comes from a test platform for full-duplex communication systems on GitHub, which contains received signals and self-interference signals. This dataset is used to train CZT-ML-CVNN and CLDNN models. The data includes time series, spectrum information, and annotations of received signals.

[0195] b. Network training:

[0196] The Keras library is used to train the CZT-ML-CVNN and CLDNN networks. The goal is to optimize the model by minimizing the error between the predicted interference signal and the actual interference signal. and the error between the real interference signal r(n).

[0197] Mini-batch gradient descent or other optimization algorithms are used during training.

[0198] c. Real-time deployment:

[0199] Xilinx Vitis AI is an AI inference development platform designed for Xilinx FPGA and ACAP (Adaptive Compute Acceleration Platform) devices. It combines hardware acceleration and a flexible software tool chain to provide a solution for the efficient deployment of deep learning models in edge devices and data centers. It supports models exported from mainstream deep learning frameworks such as TensorFlow and PyTorch. Select a supported FPGA chip (such as XCVU13P), build a software environment, deploy the trained CZT-ML-CVNN and CLDNN models to the FPGA, and process the received signal in real time.

[0200] d. Performance evaluation:

[0201] Compare the error between the output signal and the ideal signal, and calculate performance indicators such as bit error rate (BER) and signal-to-noise ratio (SNR) improvement.

[0202] The hardware implementation advantages of the embodiments of the present invention are as follows:

[0203] Use Xilinx Vitis AI to efficiently deploy deep learning models to FPGA platforms;

[0204] Strong real-time performance: FPGA’s low latency makes it suitable for real-time processing scenarios with self-interference suppression;

[0205] Low power consumption: Compared with GPU, FPGA has a significant energy efficiency ratio.

[0206] Figure 6 This is a structural block diagram of a self-interference suppression device for full-duplex communication provided by an embodiment of the present invention, the device comprising:

[0207] An acquisition module is used to acquire input signals and perform preprocessing;

[0208] The conversion module is used to perform CZT frequency domain conversion on the input signal and output the complex spectrum signal c m (n);

[0209] Extraction module, used to convert the complex spectrum signal c m (n) Input ML-CVNN multi-layer complex-valued neural network to extract channel characteristics

[0210] Prediction module, used to transform channel characteristics Input the CLDNN network to perform channel prediction and obtain the predicted self-interference signal;

[0211] Output module, used to subtract the predicted self-interference signal from the received signal and output the received signal r out (n).

[0212] Figure 6 The functions of each module in the embodiment correspond to the contents in the corresponding method embodiment and will not be repeated here.

[0213] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A method for suppressing self-interference in full-duplex communication, characterized in that: The following steps are involved: S1: Get input signal and preprocess it; S2: Perform CZT frequency domain conversion on the input signal and output a complex spectrum signal c m (n); S3: The complex spectrum signal c m (n) Input ML-CVNN multi-layer complex-valued neural network to extract channel characteristics S4: The channel characteristics Input the CLDNN network to perform channel prediction and obtain the predicted self-interference signal; S5: Subtract the predicted self-interference signal from the received signal and output the received signal r out (n).

2. The self-interference suppression method according to claim 1, characterized in that: In step S1, the input signal includes: a baseband signal x(t) generated by a transmitting end and a received signal r(t) containing a self-interference component; the preprocessing is to digitize x(t) and r(t) to obtain equivalent discrete baseband signals x(n) and r(n).

3. The self-interference suppression method according to claim 2, characterized in that: The step S2 specifically includes the following steps: S2.1: CZT calculation of Doppler spectrum: Use the CZT conversion formula to convert the input signal r(n) into a frequency domain signal: S2.2: Channel parameter extraction: The key frequency domain features are obtained through peak detection as follows: a m : the amplitude of each path; f m : Doppler frequency; φ m : phase shift; S2.3: Output result: Output the complex spectrum signal c m (n).

4. The self-interference suppression method according to claim 1, characterized in that: In step S3, the ML-CVNN multi-layer complex-valued neural network extraction process is as follows: Input layer: The complex spectrum signal c output by CZT m (n) as input to ML-CVNN; Hidden layer: Use the ML-CVNN multi-layer complex-valued neural network for channel modeling: In the formula, |w lji ∣: the magnitude of the connection weight; θ lji : the phase of the connection weight; Activation function: Output layer: The output layer generates the predicted channel characteristics 5. The self-interference suppression method according to claim 1, characterized in that: In step S4, the CLDNN network prediction method is as follows: Input layer: The channel characteristics extracted by the ML-CVNN multi-layer complex-valued neural network Input into the CLDNN network; CNN module: uses convolutional layers to extract frequency domain features; LSTM module: uses long short-term memory network to extract time-related features; DNN module: maps features to independent space and generates interference signal prediction results; Output layer: outputs the predicted self-interference signal.

6. The self-interference suppression method according to claim 1, characterized in that: The step S5 specifically comprises the following steps: S5.1: Signal reconstruction: Reconstructing the self-interference signal according to the predicted self-interference signal S5.2: Interference cancellation: Subtract the self-interference signal from the received signal r(n) S5.3: Output the received signal r after self-interference suppression out (n).

7. The self-interference suppression method according to claim 1, characterized in that: The hardware implementation steps of the self-interference suppression method are as follows: a. Get the dataset; b. Training the ML-CVNN multilayer complex-valued neural network and the CLDNN network; c. Deploy the trained ML-CVNN multi-layer complex-valued neural network and the CLDNN network on the FPGA to process the received signal in real time; d. Compare the error between the output signal and the ideal signal and calculate the performance index.

8. The self-interference suppression method according to claim 7, characterized in that: The mini-batch gradient descent algorithm is used during the training of the ML-CVNN multi-layer complex-valued neural network and the CLDNN network.

9. The self-interference suppression method according to claim 7, characterized in that: The data set comes from a test platform of a full-duplex communication system, and includes a received signal and a self-interference signal. The data in the data set includes a time series, spectrum information, and an annotation of the received signal.

10. A self-interference suppression device for full-duplex communication, characterized in that: The method for suppressing self-interference according to any one of claims 1 to 9 is adopted, comprising: An acquisition module is used to acquire input signals and perform preprocessing; A conversion module is used to perform CZT frequency domain conversion on the input signal and output a complex spectrum signal c m (n); An extraction module is used to extract the complex spectrum signal c m (n) Input ML-CVNN multi-layer complex-valued neural network to extract channel characteristics A prediction module is used to convert the channel characteristics Input the CLDNN network to perform channel prediction and obtain the predicted self-interference signal; An output module is used to subtract the predicted self-interference signal from the received signal and output the received signal r out (n).

Images